This is a linkpost for pacing-ai.com. I encourage folks to read it via the site (which has interactive graphs and expandables), but we’ve copied the text here as well.
We offer a concrete proposal for pacing the frontier.
2. Reallocation: Shift compute away from capabilities R&D
The first component is Verified Compute Allocation Disclosures (VCAD): AI companies regularly disclose their compute usage, and an independent auditor verifies the disclosures. The second component is to reallocate compute away from capabilities R&D and toward other use cases, such as serving customers and safety research (non-capabilities R&D).
We roughly estimate that pacing via compute reallocation starting January 2027, with a 15% cap, would delay the intelligence explosion by a year. This is most likely insufficient to prevent AI takeover (or mitigate other problems created by RSI), and so this year should be used to implement a more durable policy regime such as Plan A or Plan S.
Nearly every approach to pacing, including ones built on safety and security requirements, strongly benefits from VCAD:
Any pacing proposal that actually slows down AI development must do so either by reducing one of the inputs to AI development (such as training compute, experiment compute, or automated or human labor), or by implementing a safeguard that results in using inputs less efficiently (e.g., avoiding neuralese).
Something like VCAD is evidently needed for compute restrictions. For other input restrictions and safeguards, VCAD is highly desirable to prevent attacks of the form “don’t follow this restriction/safeguard on some compute, and hide or misreport that compute.”
With VCAD, competitors and the public can trust a company that says it is pacing. As evidence of risk grows, this could enable a “race to the top”. If regulation later mandates pacing, a VCAD process that companies have already built and stress-tested makes that regulation faster to write and more workable.
Given the urgency of the situation, we elected to publish this relatively quickly. We are not very confident in any given detail of our proposal. Nevertheless, we think it represents a step forward in concreteness for pacing proposals and hope that we and others can quickly turn this into something workable.
VCAD Implementation Proposal
We propose that AI companies take the following steps to estimate and disclose their compute usage:
How AI companies could estimate compute usage
The basic proposal is:
Step 1: Logging. Implement logging infrastructure that logs all workloads running inside the AI company, along with relevant metadata.
Each company should create a database that accounts for all the compute (measured in GPU-hours) that the company uses (more detail on the scope here). Each entry in the table should have the GPU-hours used by the workload, and depending on which of the following categories of workload it fits into, it should have the following metadata attached to the log:
Step 2: Classification.
Classify the logged jobs according to whether they contributed to capabilities R&D (detailed proposal here).
Weight the jobs by how many quality-adjusted GPU-hours they used (detailed proposal here).
Step 2 involves three substages: 2.1) Aggregate GPU-hours.2.2) Classify workloads into capabilities vs. non-capabilities.2.3) Estimate idle compute.
2.1) Aggregate GPU-hours. Getting an overall estimate of compute use requires aggregating across heterogeneous GPUs into a single unified GPU-hr metric (e.g., “H100-equivalents”). We discuss this in more detail in an appendix.
2.2) Classify workloads into capabilities vs. non-capabilities, using a mixture of AI classification and manual human classification. By “capabilities” compute, we mean compute that has any positive capabilities effects — including work that has capabilities and safety externalities. This centrally includes:
Experiments to improve algorithms and data
Data generation or filtering
Training more capable AIs
Internal inference in service of the above
Unfortunately, it’s hard to give a crisp definition of this, and there are many edge cases. The exact definition isn’t that important so long as it isn’t possible to use internal non-capabilities compute in a way that has large capabilities externalities. We discuss this in more detail in a supplement here.
2.3) Estimate idle compute. By idle compute, we mean compute that is not turned on or is not in use. This includes:
Hot spares
Downtime
New racks that haven’t started use
When Step 2 is complete, the company will have an internal estimate of compute allocation such as the following:
The categories are defined as follows:
Serving customers: compute used to serve customers, including via API, web interfaces (e.g., Claude.ai), and desktop (e.g., Claude Code). We also call this category “External Inference”.
The “trained on” box is the subset of external inference that feeds back into capabilities, either by being directly trained on, paraphrased and then trained on, or similar.
Safeguard classifiers: monitors checking for misuse
Internal Inference: inference compute used for internal AI company use. This includes evaluations, internal coding agents, and RL rollouts.
We subdivide internal inference into capabilities and safety, where safety compute refers to the internal inference that doesn’t have significant capabilities externalities.
Non-Capabilities R&D: R&D compute that has minimal capabilities externalities.
Capabilities R&D: all other R&D compute.
See Supplement C for more detail on the definitions.
Step 3: Disclosure. Publicly report a breakdown of how much compute was spent in each category, and whether it counted as capabilities (red) or non-capabilities.
The AI company publishes a disclosure of the compute allocation. We could imagine, for example, a disclosure every two weeks covering compute allocation over that period. This report should include:
An overall estimate of compute allocation
An overview of the methodology used to make and audit that estimate
An analysis of possible mistakes, discrepancies the auditor was not able to resolve, and an estimate of the error rate
Descriptions of N (e.g., 30) non-capabilities projects, randomly sampled in proportion to compute
We recommend that AI companies publish this disclosure through a channel that punishes misreporting. As a starting point, it should be part of the company’s SB 53 frontier AI framework. As soon as possible, it should move to a channel with real liability, such as being included in financial disclosures or directly required by regulation with sufficient teeth to deter violation.
Even with improved reporting channels, there would still be risk of:
Deception from AI executives
Goodharting: for the attacks we’re most worried about, it can happen without explicit deception, particularly attack 5 (external compute) and attack 7 (misclassifying capabilities compute). We’d recommend these reports explicitly make claims about the company not doing activities in these buckets.
Internal conspiracies the leadership doesn’t know about.
Despite this, such a requirement would probably galvanize a lot of self-verification internally, because the CEO and board may want to be really sure they aren’t lying in these disclosures.
Third-party auditors (3PAs) would verify the integrity of this process. We give a baseline auditing process and a robust auditing process, and estimate the error rate for each.
Baseline Auditing
Baseline auditing involves checking (and potentially reproducing parts of) the accounting process, via methods including: financial auditing, spot-checking random workloads, and re-running an independent classification pipeline.
The baseline auditing process:
Could be implemented today (it does not rely on new science).
Is sufficient to give an accurate estimate of compute allocation within AI companies today.
Is not sufficient to withstand adversarial pressure from a company trying to actively lie in their disclosure, which could happen under a pacing regime where companies are supposed to significantly limit compute used for capabilities.
It involves the following substages:
3.1 Completeness. Does the table include all the compute?
Check financial data to make sure that the total chip-hours being reported by the company match the amount of compute it should have, given its compute contracts.
Create a database of all datacenters used by the AI company with capacity information.
3.2 Database Integrity. Is the information in the table a correct accounting of workloads in the company?
Spot-check a random set of workloads to confirm that the metadata is consistent with the reported GPU-hrs, by estimating the compute requirements from the auditors’ best understanding of the stack. Optionally, this could be assisted by increased visibility into the compilers and performance of the live stack on different workloads.
Directly review the logging code to check for bugs and/or backdoors.
Check whether auditor-initiated workloads are logged correctly.
3.3 Classification. Is the mapping from the database to capabilities / non-capabilities being done correctly?[1]
Run an independent classification pipeline.
Manually review a fraction of the jobs classified by the company.
Write public summaries of non-capabilities R&D projects.
Check whether auditor-initiated workloads are classified correctly.
3.4 Review for warning signs. Are there major red flags?
Check for major contracts that subsidize inference for other large companies.
Check the distribution of non-capabilities compute. Is a single type of project absorbing huge amounts of compute?
Conduct private interviews with researchers to ask if they have observed anything suspicious.
After completing the audit, the auditor would write up a report containing:
whether they agree with the AI company's disclosure and, if not, their independent estimate of compute usage.
a description of their methodology.
limitations of their audit, along with an estimate of total error (the capabilities-equivalent compute the audit could be missing, as a percentage of overall compute).
a summary of N (e.g., 30) projects sampled from the non-capabilities bucket, weighted by compute use, with enough information to show an outside observer that each project was genuinely not capabilities work.
Attacks & Loopholes
If all we want is transparency, we think that baseline auditing would most likely be sufficient. However, if there is also a limit set on capabilities R&D compute — as we think there should be — then the stricter the limit, the greater the incentive to cheat. Baseline auditing would be insufficient to reliably verify pacing.
We think that if frontier AI companies were legally required to (i) disclose how they use their compute and (ii) use no more than X% on capabilities R&D, then the most likely outcome would be that all five leading U.S. companies comply. This is for the same reasons that they normally comply with laws — too many possible whistleblowers or other ways they could get caught. This is especially true since the point of noncompliance would be to get an advantage in AI R&D progress, but that advantage would be hard to hide — if your AI models leapfrog over those of your competitors, for example, that would raise suspicions. That being said, we think that cheating is a serious possibility, especially the kind where you do it in a plausibly deniable way so that if you get caught you don’t go straight to jail.
We lay out seven types of attack a malicious AI company could use to cheat the system, and describe two cheating scenarios — one in which they stick to plausibly deniable mechanisms (Scenario 1), and one in which there is deliberate deception by a leadership-backed conspiracy (Scenario 2). We lay out a baseline and a robust auditing proposal: baseline auditing is only effective against the first scenario, while robust auditing works against the second.
This table gives an estimate of the total error in each scenario. These are defined and justified in the Attacks and Loopholes supplement.
In practice, we recommend that regulators implement regulation as soon as possible, with provisions that phase in progressively stronger auditing regimes, to get the incentives right. AI companies should know that soon there will be auditors checking for compliance, and that later there will be stronger, savvier auditors checking harder. This builds trust in the whole system and makes cheating much less likely to be attempted in the first place.
We show the attacks below. Click on an attack to see our estimated error margin from it under each scenario.
Robust auditing adds stronger measures to withstand significant adversarial pressure, such as:
KYC + auditing for major external users
AI R&D classifiers
Auditor-managed logging infrastructure
We believe that robust auditing is sufficient to keep error to ~6% of total compute, even assuming a leadership backed conspiracy.
We propose a set of additional measures that could improve on baseline auditing.
With the additional measures proposed here, we could limit error to ~6%, assuming high-context auditors, even under Scenario 2 (leadership-backed conspiracy). Most likely not all of these measures are needed, and they could be added incrementally over time as more robustness is required from the auditing methodology and potential issues are noticed.
There are several ways our proposal could be implemented and evolve over time. Some of the ways it can vary include the number of companies it applies to, whether it is voluntary or required by regulation, the level of granularity in disclosures, and the level of verification in place.
The pilot we propose is voluntary disclosures with third-party auditing. AI companies voluntarily disclose their compute allocations per our proposal and allow auditors to immediately carry out baseline auditing and then work up to robust auditing.
Why would companies voluntarily pace?
AI companies are in a collective action problem: they are proceeding faster than they believe optimal. VCAD supplies two elements that can help resolve the problem, prompting a race to the top.
Trust: Claims of costly restraint today are not always believed by competitors. VCAD provides a way in which companies can take commercially costly actions in the interest of safety or security that competitors can trust.
Public credit: VCAD’s public disclosures (the “D”) offer a simple, commensurable benchmark for cross-company comparisons, which can help channel public and policymaker concern into a concrete request: less compute on R&D.
Interest in pacing is high and will grow with evidence of risk. Demand for reduced R&D allocations gives companies a self-interested reason to reallocate and rewards those that do, spurring a race to the top. Each reallocation would be a unilateral response to, among other things, public demand based on a public benchmark. The dynamic does not depend on, and should not involve, agreement among companies on allocation levels.
There are two additional pathways that might be alternatives if the former fails (or that can be pursued in parallel) which we find promising, which vary in how much company and political will is required:
With less company will: voluntary binding disclosures without auditing. In this scenario, the companies could still voluntarily disclose their compute allocations via financial disclosures or Congress (where lying is a criminal offense), or via the SB 53 Frontier Safety Framework (where lying carries small civil penalties).[2]
With more political will: state or federal regulation. Domestic regulation, either state law or federal legislation passed by Congress, that applies to the top 5 AI companies or more, with verification through either more binding disclosures or, better, third-party or government auditing.
Implementation over time
1. Adoption
Voluntary — Companies voluntarily disclose their compute allocations and use them to credibly demonstrate pacing.
Domestic regulation — State or federal. Options include: (i) just disclosure + verification; (ii) disclosure + verification + empowering an agency to mandate reallocations; (iii) a component of a conditional regime (see 4 below).
Long term: International regulation
These steps can happen in sequence; alternatively, the voluntary regime can be skipped or cut short by domestic regulation.
2. Coverage
US Top 2 (e.g., OpenAI and Anthropic) — These companies have a ~3-month lead over everyone else.
US Top 5 (OpenAI, Anthropic, Meta, GDM, xAI) — These companies have a ~6-month lead over everyone else.
All domestic US AI companies
Long term: International — Eventually, AI regulation that would prevent the development of misaligned superintelligence needs to be international, including in China.
3. Disclosure granularity
We think the categories of compute per our diagram in this proposal, together with some transparency on the workloads that get classified as non-capabilities, should be the baseline for what all companies do in common.
Our initial diagram + Safety Transparency — Report compute allocations for each of: capabilities R&D, non-capabilities R&D, serving customers, classifiers, and idle compute. For non-capabilities R&D, take a representative sample of that compute and publish a summary.
Granular reporting — Everything above, plus more granular categories, such as a breakdown of inference into RL rollouts/synthetic data, autonomous agent swarms, model evaluations, and employee use.
Long term: Total Research Transparency — As much as possible is public; auditors on the inside work to ensure that nearly all risk-relevant information is public, including compute allocations.
4. Combination with safety and security requirements
We don’t currently have a science of AI risk sufficient to make conditional triggers that will be robust in the future, so we think an unconditional implementation of this proposal is needed as a stopgap.
Unconditional — Set a capabilities compute cap unconditionally, and adjust it over time as needed without explicit triggers.
Long term: Conditional — A component of a mature governance regime, with safety and security requirements as triggers.
5. Verification
To be effective, disclosures must be trusted within a reasonable margin of error (e.g., <5% of total compute). The burden of verification will increase with how strong the incentives are to misreport and which actors are involved.
No external verification: SB 53 FSF — Part of the framework that AI companies are required to follow under California’s SB 53. Violation incurs a small fine (<=$1M).
External verification: Third-party embedded baseline audit — Third-party auditor gets employee-level access, allowing them to directly inspect logging infrastructure, submit jobs, and red-team the classifiers.
Third-party robust audit — The third party can do everything in our Robust Auditing section.
Long term: Hardware Verification — Robust hardware security properties, paired with well-developed software measures such as ZKPs, etc.
Frequently Asked Questions
Why pace?
Timelines to recursive self-improvement (RSI) are most likely short. Our best guess is that, absent a deliberate slowdown, AIs will fully automate the process of AI R&D before EOY 2027.
We aren’t anywhere near ready to do full RSI. This is true for many reasons.
Loss of control: Current AIs are misaligned. Continuing to improve capabilities without substantial improvements to alignment would pose a large risk of AI takeover. Even if we think current AIs are safe, recursive self-improvement will increase the pace of AI progress and reduce human understanding, so misalignment will probably get worse. We will probably need to slow down during the intelligence explosion, so we need to establish the mechanism to do so as soon as possible.
Concentration of power: Even if AI companies manage to align their “army of geniuses in the datacenters”, that would be an unacceptable concentration of power. A tiny group of CEOs or politicians could end up choosing the values of the army of geniuses which obsolete humanity. AI-powered oligarchy or dictatorship are live possibilities.
Misuse risk: Bioweapons are extremely hard to defend against. The cost and expertise needed to make an engineered pathogen are decreasing rapidly. Resilience and defense are not keeping pace.
Pacing would delay RSI. It would give us more time for:
Alignment and control R&D.
The public and governments to wake up and learn about what’s going on.
Legislation to be crafted, voted on, and implemented.
How much would compute reallocation slow things down?
This depends on what the capabilities R&D dial is set to, how much error there is in the regime, and what the production function for AI progress is with respect to R&D compute. If we assume the capabilities allocation is set to 20% and 6% error is exploited, then the resulting 26% allocation would be an immediate ~3x reduction in capabilities R&D compute for Anthropic (which uses an estimated ~75% of its compute on capabilities R&D today).
We don’t know how much overall AI progress would slow down due to an ongoing 3x reduction in R&D compute, because there are other inputs to AI progress and we don’t know how important each of them is. However, our best guess is that it would slow progress down by around 1.5-2x[3] in the short term. Over time, the pace of progress is determined by the growth rate of compute, so assuming the company’s compute buildout is similar under pacing, its speed would converge back to the previous rate.
Imagine a phantom copy of the company that has 3x less compute, growing at the same rate. This phantom company would follow behind the leader with a fixed calendar time delay, but move at the same pace (because it has as much compute as the frontier company had a fixed time ago).[4] The one-time 3x reduction in compute puts the frontier company in the phantom company’s position from that point onward, so its progress converges towards the phantom company’s trajectory.
As a result, later capability milestones are delayed more (the delay is still growing towards the fixed calendar lag), but the slowdown in the rate at which milestones are crossed shrinks towards zero. So the compute reduction mainly pushes back timelines, and its effect on the speed of takeoff is smaller the further away takeoff is. This assumes compute growth is constant; if compute growth slows in the future (as is likely, absent large AI-driven speedups), then takeoff could also be slower. In practice, though, timelines to highly capable AI might be short enough that this convergence doesn’t have time to matter much.
In a run of our AI Futures Model (which, without pacing, reaches Superhuman AI Researcher (SAR) in August 2027 and Top-Expert-Dominating-AI (TED-AI) in October 2027), a 3x cut from January 2027 delays SAR by about 4 months and TED-AI by about 5 months, and the paced company crosses both about 24% slower than it would without pacing.[5][6]
What about China?
The US should care about both the absolute level of risk from continued AI development and its relative lead over China. Even if you care only about the relative lead over China and don’t believe in the risks, the downsides of this proposal are likely much smaller than you might expect. This rests on the following core arguments:
Security is poor enough such that China could catch up to the US at any time anyway. AI companies are not on track to be secure against efforts by China to steal their IP (OC4 or OC5 efforts as defined by RAND). Therefore, assuming China decides to put effort into stealing from US AI companies, there is not much upside to continuing to develop new algorithmic secrets and models.
Slowing down may make it possible to improve security without giving China the opportunity to steal higher-risk IP in the meantime. Improving security could make it harder for China to steal AIs and training recipes at the US frontier. It might be possible to make significant security improvements for model weights, dataset and distillation (and potentially on code and algorithm confidentiality) that enable the US to maintain a more robust lead in the future. The tractability of doing so is uncertain, especially for algorithm confidentiality, but at the very least a slowdown would mean that China can steal less capable models than it otherwise would.
Slowing down buys more time to possibly reach an agreement with China. By slowing down, the US can signal seriousness about the risks of further AI development, and negotiate an agreement with China for both countries to proceed with sufficient caution, in a commensurable and verifiable way.
The US has a large compute advantage over China, so in worlds where China does not aggressively steal from the US, the US labs could carry out significant pacing while still maintaining a lead. In worlds where China isn’t aggressively stealing progress from the US, the US could adopt a ‘China-dominating’ pacing proposal in which US companies in aggregate still have more compute going to capabilities R&D than China (e.g., if the gap over Chinese companies is ~10x, a 20% compute cap still leaves US companies with at least twice as much capabilities R&D compute). This would allow the US to pace while still maintaining a lead. If China aggressively steals US progress, it would become much less clear if this strategy is sufficient to maintain a lead.
This proposal would be net negative only if both of the following are true:
The risks from continued AI development are low in the near future.
China is unlikely to put much effort into stealing US progress in the near future.
If these are both true, then the US would be giving up some of its lead for little gain. If only (1) is true, then the US isn’t really giving up its lead through this proposal (China would have caught up regardless via theft), and if only (2) is true, then spending some of the US lead was worth it because it reduced risk.
Why not a different mechanism for pacing?
We can think of the options for slowing down AI progress by looking at the set of inputs to AI progress, which are R&D compute and labor, where R&D compute can be used on experiments, training runs and AI automation.
Each of these inputs could be limited in quantity (e.g., a cap on compute spend) or regulated by safety standards (e.g., banning training runs that use neuralese algorithms). The decision to restrict inputs or require safeguards could be part of a pacing regime that is triggered by reaching a pre-registered risk level (e.g., safety-case-based pacing), reaching a pre-registered capability level (e.g., RSPs), or one that has no pre-registered triggers.
Writing good safety standards that sufficiently address the sources of risk will require a lot of institutional capacity, increased scientific understanding, and ongoing iteration, so this approach cannot be deployed quickly as a pacing measure.
Therefore, limiting inputs to AI progress is the more promising near-term pacing solution. However, limits that directly gate AI capabilities (without also gating algorithmic progress) could create an algorithmic overhang.
If we slow down via anything but limiting algorithmic progress, this creates an algorithmic overhang, which, once training resumes, results in a rapid return to something similar to the original trend (though it’s somewhat lower due to reduced AI R&D uplift). This algorithmic overhang could, in principle, become very large. Algorithmic progress is fast and is likely to accelerate in the future: estimates of algorithmic progress today range from 3x to 30x per year, and it seems plausible that algorithmic progress recently has been even faster. Because algorithmic progress is more rapid than compute progress (which is ~5x/yr), and all of the built-up progress could be immediately applied to a new training run while applying a compute overhang would be intermediated by R&D, an algorithmic overhang could be much worse than a compute overhang.
The algorithmic overhang creates two major problems:
Discontinuous rapid progress. If the capabilities cap is temporary, then lifting it would cause a discontinuous jump in capabilities, which could be very dangerous.
Poor AI company security means that the unused algorithms still accelerate trailing actors. AI companies are not robust to algorithmic leakage. It is likely that any algorithmic progress made, even if it isn't used in a final training run at the compliant AI company, could get used by other AI projects.
In an emergency situation, which we may soon find ourselves in, direct AI capability restrictions may be necessary. However, we’d recommend pairing these with measures to limit algorithmic progress, in order to prevent an algorithmic progress overhang.
There are several options for limiting algorithmic progress that we think are promising, including:
Compute allocation caps to R&D. This is the proposal we analyzed in this post.
Creating a lag in the internal deployment of AI models. This could involve creating a delay between the time that a new model is trained and when it is allowed to be deployed internally. We discuss the implementation of the deployment lag proposal in more detail here.
Researcher holidays. This could involve AI researchers only working one or two days a week, with all their associated work (e.g., experiments) also stopping on the days they don’t work.
We specifically analyzed the compute allocation cap in this post, but a similar verification methodology would be required for the other measures for limiting algorithmic progress inputs. This is because the problem will continue to have the structure where:
The company has compute that it could use for R&D but is not supposed to (up to some limit, for some categories of uses).
The company could have access to other sources of compute that it can use freely.
Supplements
Supplement A: Attacks & Loopholes — We argue the baseline methodology is not sufficiently robust to verify an AI slowdown, which would create significant adversarial pressure.
Supplement B: Robust Auditing — We argue that with additional measures, we could limit error to ~6%, even under significant adversarial pressure.
Supplement D: RSI Deployment Lags — We give a brief overview of how VCAD could be used to verify deployment lags and the differences between deployment lags and compute reallocation.
Appendices: Aggregating heterogeneous chip data; Classifying ZDR Compute; How frontier-only AI company pacing could be scoped.
We’re especially interested in a low false-negative rate, i.e., that the capabilities classifier doesn’t miss substantial fractions of the capabilities research.
There would still be a risk of plausibly deniable deception (or willful deception) from AI companies, but the scope of declarations could be expanded to try and reduce this surface area. We are unsure how much error there would be in this scenario, but it seems plausible that it would galvanize enough self-verification for <5% error, because the CEO and board may want to be really sure they aren’t lying in these disclosures.
Assuming a simple model of AI progress (Cobb-Douglas in compute and labor: C^α · L^(1−α)). With a compute share α of 0.4–0.6, a 3x cut gives 3^α ≈ 1.55x–1.9x; α = 0.5 gives 1.73x. The bounds are 1x (compute doesn’t matter) and 3x (compute is the only bottleneck). The truth seems very likely to be somewhere in the middle.
Modified AI Futures Model: The AI Futures Model (November 2026 code) with Daniel Kokotajlo’s August 2026 median parameters and these parameter changes: serial coding uplift 7x in August 2026 (preset 2x); excess uplift doubling time 3.5 months (preset 5); coding labor substitutability ρ = −0.2 (preset −2); experiment capacity ceiling from infinite coding labor 60x (preset 15x); slowdown from 10x less experiment compute 1.5x (preset 2.8x); AI research taste slope 2.3 SD per 2025-progress-year (preset 3.0); AI research taste at −1.5 SD on November 1, 2026. Trajectory without pacing.
Public AI Futures Model: The AI Futures Model (same code) with Eli Lifland’s August 2026 median parameters and these parameter changes: serial coding uplift 12x in October 2026 (preset 1.7x); excess uplift doubling time 3.5 months (preset 6.5), chosen so that SAR without pacing matches the modified run. Trajectory without pacing.
This is a linkpost for pacing-ai.com. I encourage folks to read it via the site (which has interactive graphs and expandables), but we’ve copied the text here as well.
We offer a concrete proposal for pacing the frontier.
1. Transparency: Verified Compute Allocation Disclosures (VCAD)
2. Reallocation: Shift compute away from capabilities R&D
The first component is Verified Compute Allocation Disclosures (VCAD): AI companies regularly disclose their compute usage, and an independent auditor verifies the disclosures. The second component is to reallocate compute away from capabilities R&D and toward other use cases, such as serving customers and safety research (non-capabilities R&D).
We roughly estimate that pacing via compute reallocation starting January 2027, with a 15% cap, would delay the intelligence explosion by a year. This is most likely insufficient to prevent AI takeover (or mitigate other problems created by RSI), and so this year should be used to implement a more durable policy regime such as Plan A or Plan S.
Nearly every approach to pacing, including ones built on safety and security requirements, strongly benefits from VCAD:
With VCAD, competitors and the public can trust a company that says it is pacing. As evidence of risk grows, this could enable a “race to the top”. If regulation later mandates pacing, a VCAD process that companies have already built and stress-tested makes that regulation faster to write and more workable.
Given the urgency of the situation, we elected to publish this relatively quickly. We are not very confident in any given detail of our proposal. Nevertheless, we think it represents a step forward in concreteness for pacing proposals and hope that we and others can quickly turn this into something workable.
VCAD Implementation Proposal
We propose that AI companies take the following steps to estimate and disclose their compute usage:
How AI companies could estimate compute usage
The basic proposal is:
Step 1: Logging. Implement logging infrastructure that logs all workloads running inside the AI company, along with relevant metadata.
Each company should create a database that accounts for all the compute (measured in GPU-hours) that the company uses (more detail on the scope here). Each entry in the table should have the GPU-hours used by the workload, and depending on which of the following categories of workload it fits into, it should have the following metadata attached to the log:
Step 2: Classification.
Step 2 involves three substages: 2.1) Aggregate GPU-hours. 2.2) Classify workloads into capabilities vs. non-capabilities. 2.3) Estimate idle compute.
2.1) Aggregate GPU-hours. Getting an overall estimate of compute use requires aggregating across heterogeneous GPUs into a single unified GPU-hr metric (e.g., “H100-equivalents”). We discuss this in more detail in an appendix.
2.2) Classify workloads into capabilities vs. non-capabilities, using a mixture of AI classification and manual human classification. By “capabilities” compute, we mean compute that has any positive capabilities effects — including work that has capabilities and safety externalities. This centrally includes:
Unfortunately, it’s hard to give a crisp definition of this, and there are many edge cases. The exact definition isn’t that important so long as it isn’t possible to use internal non-capabilities compute in a way that has large capabilities externalities. We discuss this in more detail in a supplement here.
2.3) Estimate idle compute. By idle compute, we mean compute that is not turned on or is not in use. This includes:
When Step 2 is complete, the company will have an internal estimate of compute allocation such as the following:
The categories are defined as follows:
See Supplement C for more detail on the definitions.
Step 3: Disclosure. Publicly report a breakdown of how much compute was spent in each category, and whether it counted as capabilities (red) or non-capabilities.
The AI company publishes a disclosure of the compute allocation. We could imagine, for example, a disclosure every two weeks covering compute allocation over that period. This report should include:
We recommend that AI companies publish this disclosure through a channel that punishes misreporting. As a starting point, it should be part of the company’s SB 53 frontier AI framework. As soon as possible, it should move to a channel with real liability, such as being included in financial disclosures or directly required by regulation with sufficient teeth to deter violation.
Even with improved reporting channels, there would still be risk of:
Despite this, such a requirement would probably galvanize a lot of self-verification internally, because the CEO and board may want to be really sure they aren’t lying in these disclosures.
Third-party auditors (3PAs) would verify the integrity of this process. We give a baseline auditing process and a robust auditing process, and estimate the error rate for each.
Baseline Auditing
Baseline auditing involves checking (and potentially reproducing parts of) the accounting process, via methods including: financial auditing, spot-checking random workloads, and re-running an independent classification pipeline.
The baseline auditing process:
It involves the following substages:
3.1 Completeness. Does the table include all the compute?
3.2 Database Integrity. Is the information in the table a correct accounting of workloads in the company?
3.3 Classification. Is the mapping from the database to capabilities / non-capabilities being done correctly?[1]
3.4 Review for warning signs. Are there major red flags?
After completing the audit, the auditor would write up a report containing:
Attacks & Loopholes
If all we want is transparency, we think that baseline auditing would most likely be sufficient. However, if there is also a limit set on capabilities R&D compute — as we think there should be — then the stricter the limit, the greater the incentive to cheat. Baseline auditing would be insufficient to reliably verify pacing.
We think that if frontier AI companies were legally required to (i) disclose how they use their compute and (ii) use no more than X% on capabilities R&D, then the most likely outcome would be that all five leading U.S. companies comply. This is for the same reasons that they normally comply with laws — too many possible whistleblowers or other ways they could get caught. This is especially true since the point of noncompliance would be to get an advantage in AI R&D progress, but that advantage would be hard to hide — if your AI models leapfrog over those of your competitors, for example, that would raise suspicions. That being said, we think that cheating is a serious possibility, especially the kind where you do it in a plausibly deniable way so that if you get caught you don’t go straight to jail.
We lay out seven types of attack a malicious AI company could use to cheat the system, and describe two cheating scenarios — one in which they stick to plausibly deniable mechanisms (Scenario 1), and one in which there is deliberate deception by a leadership-backed conspiracy (Scenario 2). We lay out a baseline and a robust auditing proposal: baseline auditing is only effective against the first scenario, while robust auditing works against the second.
This table gives an estimate of the total error in each scenario. These are defined and justified in the Attacks and Loopholes supplement.
In practice, we recommend that regulators implement regulation as soon as possible, with provisions that phase in progressively stronger auditing regimes, to get the incentives right. AI companies should know that soon there will be auditors checking for compliance, and that later there will be stronger, savvier auditors checking harder. This builds trust in the whole system and makes cheating much less likely to be attempted in the first place.
We show the attacks below. Click on an attack to see our estimated error margin from it under each scenario.
Read more → Supplement A · Attacks & Loopholes
Robust Auditing
Robust auditing adds stronger measures to withstand significant adversarial pressure, such as:
We believe that robust auditing is sufficient to keep error to ~6% of total compute, even assuming a leadership backed conspiracy.
We propose a set of additional measures that could improve on baseline auditing.
With the additional measures proposed here, we could limit error to ~6%, assuming high-context auditors, even under Scenario 2 (leadership-backed conspiracy). Most likely not all of these measures are needed, and they could be added incrementally over time as more robustness is required from the auditing methodology and potential issues are noticed.
Read more → Supplement B · Robust Auditing
Implementation pathways
There are several ways our proposal could be implemented and evolve over time. Some of the ways it can vary include the number of companies it applies to, whether it is voluntary or required by regulation, the level of granularity in disclosures, and the level of verification in place.
The pilot we propose is voluntary disclosures with third-party auditing. AI companies voluntarily disclose their compute allocations per our proposal and allow auditors to immediately carry out baseline auditing and then work up to robust auditing.
Why would companies voluntarily pace?
AI companies are in a collective action problem: they are proceeding faster than they believe optimal. VCAD supplies two elements that can help resolve the problem, prompting a race to the top.
Interest in pacing is high and will grow with evidence of risk. Demand for reduced R&D allocations gives companies a self-interested reason to reallocate and rewards those that do, spurring a race to the top. Each reallocation would be a unilateral response to, among other things, public demand based on a public benchmark. The dynamic does not depend on, and should not involve, agreement among companies on allocation levels.
There are two additional pathways that might be alternatives if the former fails (or that can be pursued in parallel) which we find promising, which vary in how much company and political will is required:
Implementation over time
1. Adoption
These steps can happen in sequence; alternatively, the voluntary regime can be skipped or cut short by domestic regulation.
2. Coverage
3. Disclosure granularity
We think the categories of compute per our diagram in this proposal, together with some transparency on the workloads that get classified as non-capabilities, should be the baseline for what all companies do in common.
4. Combination with safety and security requirements
We don’t currently have a science of AI risk sufficient to make conditional triggers that will be robust in the future, so we think an unconditional implementation of this proposal is needed as a stopgap.
5. Verification
To be effective, disclosures must be trusted within a reasonable margin of error (e.g., <5% of total compute). The burden of verification will increase with how strong the incentives are to misreport and which actors are involved.
Frequently Asked Questions
Why pace?
How much would compute reallocation slow things down?
This depends on what the capabilities R&D dial is set to, how much error there is in the regime, and what the production function for AI progress is with respect to R&D compute. If we assume the capabilities allocation is set to 20% and 6% error is exploited, then the resulting 26% allocation would be an immediate ~3x reduction in capabilities R&D compute for Anthropic (which uses an estimated ~75% of its compute on capabilities R&D today).
We don’t know how much overall AI progress would slow down due to an ongoing 3x reduction in R&D compute, because there are other inputs to AI progress and we don’t know how important each of them is. However, our best guess is that it would slow progress down by around 1.5-2x[3] in the short term. Over time, the pace of progress is determined by the growth rate of compute, so assuming the company’s compute buildout is similar under pacing, its speed would converge back to the previous rate.
Imagine a phantom copy of the company that has 3x less compute, growing at the same rate. This phantom company would follow behind the leader with a fixed calendar time delay, but move at the same pace (because it has as much compute as the frontier company had a fixed time ago).[4] The one-time 3x reduction in compute puts the frontier company in the phantom company’s position from that point onward, so its progress converges towards the phantom company’s trajectory.
As a result, later capability milestones are delayed more (the delay is still growing towards the fixed calendar lag), but the slowdown in the rate at which milestones are crossed shrinks towards zero. So the compute reduction mainly pushes back timelines, and its effect on the speed of takeoff is smaller the further away takeoff is. This assumes compute growth is constant; if compute growth slows in the future (as is likely, absent large AI-driven speedups), then takeoff could also be slower. In practice, though, timelines to highly capable AI might be short enough that this convergence doesn’t have time to matter much.
In a run of our AI Futures Model (which, without pacing, reaches Superhuman AI Researcher (SAR) in August 2027 and Top-Expert-Dominating-AI (TED-AI) in October 2027), a 3x cut from January 2027 delays SAR by about 4 months and TED-AI by about 5 months, and the paced company crosses both about 24% slower than it would without pacing.[5][6]
What about China?
The US should care about both the absolute level of risk from continued AI development and its relative lead over China. Even if you care only about the relative lead over China and don’t believe in the risks, the downsides of this proposal are likely much smaller than you might expect. This rests on the following core arguments:
AI companies are not on track to be secure against efforts by China to steal their IP (OC4 or OC5 efforts as defined by RAND). Therefore, assuming China decides to put effort into stealing from US AI companies, there is not much upside to continuing to develop new algorithmic secrets and models.
Improving security could make it harder for China to steal AIs and training recipes at the US frontier. It might be possible to make significant security improvements for model weights, dataset and distillation (and potentially on code and algorithm confidentiality) that enable the US to maintain a more robust lead in the future. The tractability of doing so is uncertain, especially for algorithm confidentiality, but at the very least a slowdown would mean that China can steal less capable models than it otherwise would.
By slowing down, the US can signal seriousness about the risks of further AI development, and negotiate an agreement with China for both countries to proceed with sufficient caution, in a commensurable and verifiable way.
In worlds where China isn’t aggressively stealing progress from the US, the US could adopt a ‘China-dominating’ pacing proposal in which US companies in aggregate still have more compute going to capabilities R&D than China (e.g., if the gap over Chinese companies is ~10x, a 20% compute cap still leaves US companies with at least twice as much capabilities R&D compute). This would allow the US to pace while still maintaining a lead. If China aggressively steals US progress, it would become much less clear if this strategy is sufficient to maintain a lead.
This proposal would be net negative only if both of the following are true:
If these are both true, then the US would be giving up some of its lead for little gain. If only (1) is true, then the US isn’t really giving up its lead through this proposal (China would have caught up regardless via theft), and if only (2) is true, then spending some of the US lead was worth it because it reduced risk.
Why not a different mechanism for pacing?
We can think of the options for slowing down AI progress by looking at the set of inputs to AI progress, which are R&D compute and labor, where R&D compute can be used on experiments, training runs and AI automation.
Each of these inputs could be limited in quantity (e.g., a cap on compute spend) or regulated by safety standards (e.g., banning training runs that use neuralese algorithms). The decision to restrict inputs or require safeguards could be part of a pacing regime that is triggered by reaching a pre-registered risk level (e.g., safety-case-based pacing), reaching a pre-registered capability level (e.g., RSPs), or one that has no pre-registered triggers.
Writing good safety standards that sufficiently address the sources of risk will require a lot of institutional capacity, increased scientific understanding, and ongoing iteration, so this approach cannot be deployed quickly as a pacing measure.
Therefore, limiting inputs to AI progress is the more promising near-term pacing solution. However, limits that directly gate AI capabilities (without also gating algorithmic progress) could create an algorithmic overhang.
If we slow down via anything but limiting algorithmic progress, this creates an algorithmic overhang, which, once training resumes, results in a rapid return to something similar to the original trend (though it’s somewhat lower due to reduced AI R&D uplift). This algorithmic overhang could, in principle, become very large. Algorithmic progress is fast and is likely to accelerate in the future: estimates of algorithmic progress today range from 3x to 30x per year, and it seems plausible that algorithmic progress recently has been even faster. Because algorithmic progress is more rapid than compute progress (which is ~5x/yr), and all of the built-up progress could be immediately applied to a new training run while applying a compute overhang would be intermediated by R&D, an algorithmic overhang could be much worse than a compute overhang.
The algorithmic overhang creates two major problems:
In an emergency situation, which we may soon find ourselves in, direct AI capability restrictions may be necessary. However, we’d recommend pairing these with measures to limit algorithmic progress, in order to prevent an algorithmic progress overhang.
There are several options for limiting algorithmic progress that we think are promising, including:
We specifically analyzed the compute allocation cap in this post, but a similar verification methodology would be required for the other measures for limiting algorithmic progress inputs. This is because the problem will continue to have the structure where:
Supplements
We’re especially interested in a low false-negative rate, i.e., that the capabilities classifier doesn’t miss substantial fractions of the capabilities research.
There would still be a risk of plausibly deniable deception (or willful deception) from AI companies, but the scope of declarations could be expanded to try and reduce this surface area. We are unsure how much error there would be in this scenario, but it seems plausible that it would galvanize enough self-verification for <5% error, because the CEO and board may want to be really sure they aren’t lying in these disclosures.
Assuming a simple model of AI progress (Cobb-Douglas in compute and labor: C^α · L^(1−α)). With a compute share α of 0.4–0.6, a 3x cut gives 3^α ≈ 1.55x–1.9x; α = 0.5 gives 1.73x. The bounds are 1x (compute doesn’t matter) and 3x (compute is the only bottleneck). The truth seems very likely to be somewhere in the middle.
Assuming they also have 3x less of all other inputs to AI progress.
Modified AI Futures Model: The AI Futures Model (November 2026 code) with Daniel Kokotajlo’s August 2026 median parameters and these parameter changes: serial coding uplift 7x in August 2026 (preset 2x); excess uplift doubling time 3.5 months (preset 5); coding labor substitutability ρ = −0.2 (preset −2); experiment capacity ceiling from infinite coding labor 60x (preset 15x); slowdown from 10x less experiment compute 1.5x (preset 2.8x); AI research taste slope 2.3 SD per 2025-progress-year (preset 3.0); AI research taste at −1.5 SD on November 1, 2026. Trajectory without pacing.
Public AI Futures Model: The AI Futures Model (same code) with Eli Lifland’s August 2026 median parameters and these parameter changes: serial coding uplift 12x in October 2026 (preset 1.7x); excess uplift doubling time 3.5 months (preset 6.5), chosen so that SAR without pacing matches the modified run. Trajectory without pacing.